Explainable header-centric framework for metadata-only column type annotation and data quality assessment
This paper (arXiv:2610.10541v1) presents an explainable, header-centric framework for metadata-only Column Type Annotation (CTA) and Data Quality Assessment (DQA) that maps column headers to 39 interpretable FinalFormat types using curated lexical resources and token-level traceability (SourceKeywords). Each type triggers rule-based detections of Data Quality Issues (DQIs) which are aggregated into HeadersIQ, a lightweight data source-level quality metric; the approach was evaluated on ~120,000 header columns across heterogeneous benchmarks (UCI, Prague, Kaggle, VizNet/Sato, SOTAB, T2Dv2, SemTab 2024), with modest official SemTab GT-strict results but a blinded diagnostic audit attributing many mismatches to benchmark granularity, aliasing, and ontology-selection effects.